Noise correlations in shear flows T E P

advertisement
Eur. Phys. J. B 33, 373–378 (2003)
DOI: 10.1140/epjb/e2003-00178-3
THE EUROPEAN
PHYSICAL JOURNAL B
Noise correlations in shear flows
B. Eckhardt1,a and R. Pandit2,b
1
2
Fachbereich Physik, Philipps-Universität Marburg, 35032 Marburg, Germany
Centre for Condensed Matter Theory, Department of Physics, Indian Institute of Science, Bangalore, 560 012, India
Received 26 February 2002 / Received in final form 11 March 2003
c EDP Sciences, Società Italiana di Fisica, Springer-Verlag 2003
Published online 20 June 2003 – Abstract. We consider the effects of a shear on velocity fluctuations in a flow. The shear gives rise to a
transient amplification that not only influences the amplitude of perturbations but also their time correlations. We show that, in the presence of white noise, time correlations of transversal velocity components are
exponential and that correlations of the longitudinal components are exponential with an algebraic prefactor. Cross correlations between transversal and downstream components are strongly asymmetric and
provide a clear indication of non-normal amplification. We suggest experimental tests of our predictions.
PACS. 47.27.-i Turbulent flows, convection, and heat transfer – 05.40.Ca Noise – 47.20.Ft Instability of
shear flows
1 Introduction
There has been a resurgence of theoretical interest in
shear-driven flows, ubiquitous both in nature and in the
laboratory [1], because the linearization of the NavierStokes equation about a laminar shear profile gives rise
to a non-normal linear operator. The eigenstates of this
operator are not orthogonal so, even if the laminar profile
is asymptotically stable, perturbations can grow as transients for a while before submitting eventually to viscous
damping [1–5]. This effect occurs generically if non-normal
linear operators arise during a linear-stability analysis and
it has been discussed at length in various contexts [6], including the hydrodynamical one we concentrate on here.
Several authors have advocated the use of pseudospectra in problems involving non-normal operators [4,7].
A complementary approach is to characterise the behaviour of non-normal systems by studying their response
to an externally imposed noise. Farrell and Ioannou [8,9]
showed that, in general, this results in an increased variance; they determined amplification rates and the perturbations that give rise to the largest amplitudes for shear
flows [10]. Rather little has been said about temporal correlations in such systems; an important, early exception is
the study by Onuki [11] whose emphasis is quite different
from ours as we explain below. But the temporal relation
between lift-up and other instabilities is at the heart of
the steady cycle that Waleffe et al. propose as the main
mechanism for sustaining turbulent fluctuations in shear
flows [12–14]. The idea is that downstream vortices give
rise to downstream streaks by non-normal amplification,
that the streaks go unstable to vertical vortex formation
a
e-mail: bruno.eckhardt@physik.uni-marburg.de
b
Also at: Jawaharlal Nehru Centre for Advanced Scientific
Research, Bangalore, India
and that finally these vertical vortices are folded over into
downstream vortices by the linear shear profile.
It is our aim here to present results on correlations in
linear, non-normal systems driven by white noise and to
discuss their relevance to perturbed shear flows. Specifically, we will show that temporal cross correlations between a streak and a vortex component can provide unambiguous signs of lift-up and non-normal amplification.
The formalism we use to study the fluctuations is similar to Rapid Distortion Theory or RDT [15,16]. We assume that we have a prescribed strong shear and follow
a perturbation in Fourier space, using Kelvin modes. In
order to avoid technical difficulties related to advecting
frames of reference or time-dependent wave vectors, we
use perturbations without variation in the flow direction.
Nevertheless, as we will discuss later on, there is numerical evidence that this does not significantly affect the main
conclusion that we draw about cross-correlations.
The outline of the paper is as follows: In Section 2 we
discuss a stochastic model with two degrees of freedom,
with one vortex and one shear component, that highlights
the non-normal coupling and the temporal correlations to
be expected. We then turn in Section 3 to a full discussion of temporal correlations in the stochastically forced
Navier-Stokes equation linearised about a laminar shear
profile. Section 4 contains concluding remarks, suggestions
for experimental tests of our predictions, and a discussion
of the relation of our work to earlier studies.
2 A simple model
2.1 Vortex-streak coupling in shear flows
Consider a linear shear profile, U0 = Szex. Coordinates are chosen in the meteorological convention, with x
374
The European Physical Journal B
1
wallnormal
0.5
0
−0.5
−1
0
1
2
3
spanwise
4
5
6
Fig. 1. Sketch of a vortex and a streak in a shear flow, looking downstream. The velocity components of the vortex lie in the
plane and are indicated by the arrows. The streak has only one velocity component perpendicular to the plane; its values are
indicated by a grey scale. The extrema of the streak (light and dark regions) are located near the maximal normal velocities.
All scales of length and velocity are arbitrary.
pointing downstream, y in the spanwise direction and z
pointing in the direction of the shear. The Navier-Stokes
equation for the fluid velocity u linearised around this
flow is
∂t u + (u · ∇)U0 + (U0 · ∇)u = −∇p + ν∆u,
(1)
where p is the kinematic pressure and ν the kinematic
viscosity of the fluid. To keep the analysis as simple as
possible we work with the Fourier modes appropriate for
periodic boundary conditions in spanwise and downstream
directions, and free-slip boundary conditions on two parallel planes in the normal direction. The analysis of the
linear problem with Kelvin modes [17–19] shows that modulations in the downstream direction give rise to a timedependent wave vector and faster-than-exponential damping. Farrell and Ioannou [10] also show that the most
important modes for non-normal amplification do not
have a downstream variation. Therefore, we consider only
perturbations with wave numbers k = (0, ky , kz ), where
ky is continuous and kz = πn/d, with n an integer and d
the distance between the bounding planes. These considerations help us to identify scales for the flows considered
in this section; their full significance will become clear in
the next section.
In order to highlight the essentials of non-normal amplification, we now take a velocity field consisting of two
modes, namely, a spanwise streak


α sin αy cos βz


0
(2)
us = 
,
0
and a downstream vortex

0



uω =  β cos αy sin βz  ,
(3)
−α sin αy cos βz
with amplitudes s(t) and ω(t), i.e.,
u = s(t)us + ω(t)uω .
(4)
The two modes are sketched in Figure 1. In the linearised
equation the pressure disappears as the velocity fields
are divergence-free. The term (U0 · ∇)u drops out and
(u · ∇)U0 results in a coupling between vortex and streak:
s
ṡ
−ν(α2 + β 2 )
S
·
(5)
=
2
2
ω
ω̇
0
−ν(α + β )
The matrix on the right hand side is not symmetric because of the coupling of both modes through the term
(u · ∇)U0 = Su3 ex . The dynamics that follows from the
non-normal system (5) has exponentially decaying vortices
that drive spanwise streaks: if s0 and ω0 denote the initial
amplitudes, then
2
s(t) = (s0 + Sω0 t)e−ν(α
ω(t) = ω0 e
−ν(α2 +β 2 )t
.
+β 2 )t
;
(6)
Clearly, even if there is no streak initially (i.e., s0 = 0),
there will be one as time progresses as a consequence of
the mixing induced by the downstream vortex. Eventually,
however, both will decay. The maximal amplitude of the
streak follows from the maximum of t exp(−ν(α2 + β 2 )t),
which occurs at a time
1
·
(7)
tmax =
2
ν(α + β 2 )
Since the maximal amplitude of the downstream component of the streak us is α, the maximal modulation of the
downstream velocity component follows to be
α
us,max = S
ω0 .
(8)
ν(α2 + β 2 )
Note that this is proportional to the shear (the larger the
shear the stronger the amplification) and to the inverse
of the viscosity. This model also allows us to draw conclusions about the sizes of the vortices: Assume that the
thickness in the shear direction, i.e., the wave number β,
is given. Then the maximal value of us,max is obtained
from differentiation of equation (8) for α = β. If β = π/d,
with d the separation between the plates, the wavelength
in the spanwise direction is about twice this width, i.e.,
λspan 2π/β = 2d.
B. Eckhardt and R. Pandit: Noise correlations in shear flows
2.2 Vortex-shear interactions with noise
Suppose now that we add noise to the above system.
Since the two components are already divergence-free, we
model the noise by adding the random perturbations ξs (t)
and ξω (t) to each component as follows:
ṡ
−λ S
s
ξs
·
(9)
=
+
ω̇
0 −λ
ω
ξω
For simplicity we assume that
ξω (t)ξω (t ) = Dω δ(t − t )
ξs (t)ξω (t ) = 0.
(10)
The formalism is easily expanded to more general forms.
For notational convenience and since the quantitative dependence on wave numbers and viscosity is of minor interest in this section, we have defined λ = ν(α2 + β 2 ). If
we start at t = −∞, the formal solution is
t
ω(t) =
dt e−λ(t−t ) ξω (t );
−∞
t
s(t) =
−∞
+S
dt e−λ(t−t ) ξs (t )
t
dt
−∞
t
−∞
dt e−λ(t−t ) e−λ(t −t ) ξω (t ).
(11)
With this choice of initial time any effects of initial conditions drop out and the effects of noise are highlighted.
2.3 Correlations
The equations for the vorticity with their exponential
damping and a white-noise source represent an OrnsteinUhlenbeck process and thus lead to exponentially decaying
correlations [20,21]. Formally this follows from
Cω,ω (t, τ ) = ω(t)ω(τ ) = e−λ(t+τ )
t
τ
×
dt
dτ eλ(t +τ ) ξω (t )ξω (τ ), (12)
−∞
−∞
which yields, in conjunction with equation (10) and the
restriction of the final integral up to min(t, τ ),
Cω,ω (t, τ ) =
Dω −λ|t−τ |
e
.
2λ
an exponentially decaying correlation function as in equation (13). The non-normal amplification gives rise to a
contribution from the vortices,
(vortex)
(t, τ ) =
Cs,s
t
t
dt
dt
−∞
τ
−∞
−∞
dτ τ
−∞
dτ S 2 e−λ(t−t
+τ −τ )
× ξω (t )ξω (τ )·
(14)
Again via equation (10) and the constraint on the domain
of integration to min(t , τ ), the vortex contribution to the
streak-streak correlation function becomes
ξs (t)ξs (t ) = Ds δ(t − t )
375
(vortex)
Cs,s
(t, τ ) =
The correlation function for the streaks has two components, one resulting from the noise in the streak components, and one from the non-normal amplification. The
cross terms involving the noise sources in both streak and
vortex drop out since they are uncorrelated. The streak
contribution is again an Ornstein-Uhlenbeck process with
(15)
This shows clearly an algebraic contribution that results
from the transient-growth characteristic of a non-normal
system. For small time-differences, the cusp at the origin
for the Ornstein-Uhlenbeck process becomes a rounded,
quadratic maximum. The total correlation function
Ds
S 2 Dω
Cs,s (t, τ ) =
(1
+
λ|∆|)
e−λ|∆| ,
(16)
+
2λ
4λ3
with the time difference ∆ ≡ t − τ , is symmetric in ∆
and decays monotonically. Formally, this calculation is
equivalent to that for a linearly damped process exposed
to both white noise (the driving of the streaks) and exponentially correlated noise (the cross coupling from the
vortices) [20,21]. What is specific to our case is that the
coloured-noise component is generated dynamically by
the structure of the non-normal system. Cross correlations
between vortex and streak can be calculated similarly; and
they provide interesting insights into non-normal amplification. We define
Cω,s (t, τ ) = ω(t)s(τ )·
(17)
Then, for t > τ , i.e., if the streak is probed before the
vortex, cross correlations decay exponentially as
Cω,s (t, τ ) =
S
Dω e−λ(t−τ ) ,
4λ2
for t > τ.
(18)
However, for t < τ , i.e., if the streak is probed after the
vortex, the driving of the streak by the vortex gives rise
to the correlation function
Cω,s (t, τ ) =
(13)
S 2 Dω
(1 + λ|t − τ |) e−λ|t−τ | .
4λ3
S
Dω (1 + 2λ(τ − t)) e−λ(τ −t) ,
4λ2
for t < τ,
(19)
which increases first, for short times, and then decreases.
It has a maximum that lies about 21% above the value
at zero, at λ(τ − t) = 1/2. The cross correlation thus
provides a convenient and characteristic measure of the
lift-up effect. A comparison between the three correlation
functions discussed above is shown in Figure 2.
376
The European Physical Journal B
1.2
vortex−streak
1.0
streak−streak
correlation functions
vortex−vortex
0.8
0.6
0.4
0.2
0.0
−5
−4
−3
−2
−1
0
λ∆
1
2
3
4
5
Fig. 2. Correlation functions (see text) for vortex-streak coupling with noise of equal intensity in each component. The
correlation functions are normalised by their values at zero, ∆ is the time difference, and λ, which has dimensions of inverse
time, is defined in the text.
3 Noisy shear flows
In order to model noise in a shear flow, we have to allow
for spatially fluctuating velocity fields. The dynamics we
want to study, the interaction between streaks and vortices, is mediated by the background shear flow. While
the detailed shape of either object and also the gradient
of the shear flow will depend on the specific boundary
conditions, the conclusions about the correlation function
should be less sensitive. For instance, the vortex-streak
recycling mechanism is very similar in flows with rigid
boundary conditions and in those with free-slip boundary
conditions [13,14]. We, therefore, choose the theoretically
most convenient form of free slip boundary conditions so
that we can use a Fourier representation for the velocity
field,
u(x, t) =
u(k, t)eik·x
(20)
k
and also for the noise,
ξ(x, t) =
where A(k) is the amplitude of the variance (whose precise functional form does not matter for the moment but
which will be discussed further at the end of this section),
Pij (k) = (δij −ki kj /k 2 ) is the transverse projector that enforces the incompressibility condition, and i and j denote
Cartesian components. We assume, for simplicity, that the
system is infinite; the effects of boundary conditions will
be included presently by suitable constraints on kz as in
the previous section. The temporal correlation functions
we would like to calculate (cf. Sect. 2) are
Cij (t, t ) = ui (t)uj (t ),
(25)
where · · · stands for an integration over all points in
the volume under consideration and an average over different realizations of the noise. Given our Fourier representation this becomes
Cij (t, t ) =
ui (k, t)uj (−k, t );
(26)
k
ξ(k, t)eik·x .
(21)
k
The linearized Navier-Stokes equation (1) becomes
u̇(k, t) + Suz (k, t)ex = −νk2 u(k, t) + ξ(k, t).
(22)
As in the previous section we assume that the wave vectors
have no downstream component. The stochastic force ξ (or
noise) satisfies
ξi (k, t) = 0,
ξi (k, t)ξj (k , t ) = A(k)Pij (k)δ(k + k )δ(t − t ),
(23)
(24)
now only an average over different realizations of the noise
remains for every wave number k. This average can be
calculated for each wave number individually and the sum
can then be evaluated. Let us first write equation (22) in
component form:
u̇x (k, t) + Suz (k, t) = −νk 2 ux (k, t) + ξx (k, t);
2
u̇y (k, t) = −νk uy (k, t) + ξy (k, t);
2
u̇z (k, t) = −νk uz (k, t) + ξz (k, t).
(27)
(28)
(29)
The projection onto divergence-free velocity fields and
noise terms implies that the last two equations are coupled; they can thus be represented by a single field uω from
B. Eckhardt and R. Pandit: Noise correlations in shear flows
which the y and z Cartesian components can be obtained
as follows:
uy (k, t) = −uω (k, t)kz /k;
uz (k, t) = uω (k, t)ky /k.
(30)
(31)
Now only two equations remain for each k:
u̇x (k, t) + (Sky /k)uω (k, t) = −νk 2 ux (k, t) + ξx (k, t);
(32)
u̇ω (k, t) = −νk 2 uω (k, t) + ξω (k, t);
(33)
and the noise terms satisfy
ξa (k, t)ξb (k , t ) = A(k)δab δ(k + k )δ(t − t ),
(34)
where the indices a and b stand for, respectively, the noise
components in downstream (x) and perpendicular direction (of vortex form, hence index ω). Equations (32, 33)
are of the same form as equation (9) so we can identify λ with νk 2 and S in equation (9) with S ky /k in
equation (32). We can therefore use the correlation functions (13, 16) and (18, 19) calculated previously along with
the expansion of the velocity field (25) and (26) to obtain
the correlation functions between spanwise and normalvelocity component (subscripts 2 and 3, respectively),
C2,2 =
k2
z
k
C3,3 =
ky2
k
C2,3 =
k2
k2
uω (k, t)uω (−k, t ),
(35)
uω (k, t)uω (−k, t ),
(36)
−kz ky
k
k2
uω (k, t)uω (−k, t ),
377
and
C3,1 (t, t ) =
Sky A(k)
2 (1 + 2νk 2 (t − t))e−νk (t −t) ,
νk 3 4νk 2
k
for t < t .
(41)
Interchanging 1 and 3 shows the symmetry C1,3 (t, t ) =
C3,1 (t , t).
The spatial dependence or, equivalently, the wavevector dependence of the correlation functions is given
by the summands in equations (35–37). These correlation
functions have high powers of the absolute value of k in the
denominator, so all expressions are divergent for k → 0,
and thus dominated by the smallest wave numbers. In the
absence of shear, this divergence is much milder if we use
the form of A(k) appropriate for thermal noise [11]. However, more singular forms have been used for A(k) in an
effort to obtain the Kolmogorov k −5/3 energy spectrum
at the level of a one-loop renormalization group [22]. Of
course the nonlinear term plays a crucial role in such calculations so it is not clear whether it is appropriate to use
such singular forms for A(k) in our study which does not
include nonlinearities.
So, unless the form of A(k) favors a different selection,
the correlations will be dominated by the smallest wave
numbers. Given the bounding surfaces in the z-direction,
separated by a distance d, we have a normal wave number
kz = nπ/d, with n a non-vanishing integer, and a continuous spanwise wave number ky . The smallest value of
n is n = 1 and the associated optimal ky is nπ/d. Rigid
boundary conditions may change these scales somewhat.
(37)
4 Concluding remarks
in terms of
2
1
A(k)e−νk |t−t | .
(38)
2νk 2
Thus these correlation functions decay exponentially; and
unless there is an asymmetry in the weighting of wave
numbers through the amplitude A(k), the cross correlation C2,3 vanishes on account of the antisymmetry under
reflection of each wave number individually. The correlation function for the downstream component (subscript 1)
becomes
2 2 A(k)
S
k
y
C1,1 (t, t ) =
1 + 2 6 1 + νk 2 |∆|
2νk 2
2ν k
uω (k, t)uω (−k, t ) =
k
× e−νk
2
|∆|
, (39)
where ∆ = t − t is the difference in times. The cross
correlations between vortex (as measured by the normal,
i = 3, component of the velocity) and streak (measured
by the downstream, j = 1, component) becomes
C3,1 (t, t ) =
Sky2 A(k)
k
νk 3 4νk 2
for t > t ,
e−νk
2
(t−t )
,
(40)
We have discussed correlations in noisy shear flow, both
at the level of a simple model with two degrees of freedom and in a model with spatially varying velocity fields.
The main result is that the non-normal coupling between
downstream vortices and streaks is reflected in the temporal cross correlation function between the downstream
and normal velocity components.
While there have been many studies of spatial correlations (Townsend [23] summarizes a large body of literature), few have addressed temporal correlations and
hardly any the cross correlations we described here. For
instance, Farrell et al. [8] study only the variance maintained by forcing a noisy shear flow. Onuki [11] concentrates on the case in which A(k) has the form appropriate
for thermal noise and then calculates the renormalisation
of the viscosity because of the nonlinear term. Bassam
et al. [24] concentrates on energy amplification but not
on the correlation functions we consider. Blackwelder and
Kovasznay [25] measured the cross correlations but their
diagrams do not provide sufficient detail at short times.
It is, therefore, only in recent numerical work on Lagrangian models [15] that the Lagrangian version of this
cross correlation has been considered [26]. Analysis of Lagrangian data shows indeed the asymmetry in the cross
378
The European Physical Journal B
correlation due to the non-normal effects. Eulerian cross
correlations have been studied in numerical simulations of
turbulent shear flows [27] and they also show the cross
correlations and their asymmetry in time [28].
These numerical simulations also indicate that the relation of our linear model with independent, additive and
white in time random perturbations to natural fluctuations in a turbulent flow is not as far fetched as it may
seem. Certain time scales and quantitative details will be
affected by the nonlinearities neglected in our calculation,
but if there is an overall shear the main mechanism of
non-normal growth will always be present, and that is at
the heart of the cross correlations we study here.
Many of the features of the cross correlations we discuss should be accessible experimentally. The exponential correlation in the vortices should be reflected in the
correlations of transversal, i.e., spanwise and normal, velocity components. The streaks are a property of the
downstream velocity component and the long correlations
should be visible there. Finally, the streak-vortex correlations should be noticeable in longitudinal-transversal correlations. We are not aware of experimental measurements
of such correlations. The most direct test would involve
fluctuations in laminar shear flows. Two experimental set
ups come to mind:
1. One could exploit the inevitable free-stream turbulence in a wind tunnel as a noise source. A shear gradient can then be imposed by a flat plate, above which a
linear shear gradient forms. Crossed hotwire anemometers can be used to characterise the statistics of the
upstream noise and the velocity field above the plate.
The free decay of the noise between the point of characterisation and the point of measurement should have
a negligible effect. Reynolds numbers should be low
enough so that no turbulence transitions in the shear
layer are induced.
2. There are several experiments or proposals for driving
and influencing flows in electrically conducting fluids
by electrical and magnetic fields [29–31]. Random perturbations of well defined strength could be induced
by random fluctuations in the currents. Ideally, flow
geometries that are closed, such as a Taylor-Couette
apparatus, or spatially localized, as in a plane Couette
flow [32,33] should be used. Such a setup should allow
for a good control of the noise and should avoid complications arising from the advection of perturbations.
Finally, we would like mention that there is growing
interest in turbulent transport in magnetohydrodynamics
in the presence of a background shear [34]. The techniques
developed here can be applied fruitfully to this case as
well, as we will show elsewhere [35].
of Advanced Research (Project No. 2404-2). We would also like
to thank Sriram Ramaswamy for discussions.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
This research was started while we were visiting the Institute of
Theoretical Physics, University of California at Santa Barbara,
USA, where it was supported in part by the US National Science Foundation under Grant No. PHY94-07194. For further
support BE would like to thank the Deutsche Forschungsgemeinschaft and RP the Indo-French Centre for the Promotion
33.
34.
35.
S. Grossmann, Rev. Mod. Phys. 72, 603 (2000)
M. Landahl, J. Fluid Mech. 98, 243 (1980).
L. Boberg, U. Brosa, Z. Naturforsch. A 43, 697 (1988)
N.L. Trefethen, A. Trefethen, S.C. Reddy, T.A. Driscoll,
Science 261, 578 (1992)
B.F. Farrell, P.J. Ioannou, Phys. Fluids A 5, 2298 (1993)
M.G. Neubert, H. Caswell, Ecology 78, 653 (1997)
P.J. Schmid, D.S. Henningson, Stability and transition in
shear flows (Springer, New York, 2001)
B.F. Farrell, P.J. Ioannou, Phys. Rev. Lett. 72, 1188
(1994)
P.J. Ioannou, J. Atmos. Sci. 51, 998 (1995)
B.F. Farrell, P.J. Ioannou, Phys. Fluids A 5, 1390 (1993)
A. Onuki, Phys. Lett. A 70, 31 (1979)
J.M. Hamilton, J. Kim, F. Waleffe, J. Fluid Mech. 287,
317 (1995)
F. Waleffe, Phys. Fluids 7, 3060 (1995)
F. Waleffe, Phys. Fluids 9, 883 (1997)
S. Pope, Turbulent Flows (Cambridge University Press,
Cambridge, 2000)
S. Nazarenko, N.K.-R. Kevlahan, B. Dubrulle, Physica D
139, 158 (2000)
Lord Kelvin, Philos. Mag. 23, 459 (1887)
A.D.D. Craik, W.O. Criminale, Proc. R. Soc. (London), A
406, 13 (1986)
B. Eckhardt, J. Kronjäger, K. Marzinzik, Nonnormal amplification in viscoelastic fluids, to be submitted
H. Risken, The Fokker-Planck Equation (Springer, Berlin,
1996)
C.W. Gardiner, Handbook of Stochastic Methods (Springer,
Berlin, 1996)
See, e.g., A. Sain, Manu, R. Pandit, Phys. Rev. Lett. 81,
4377 (1998) and references therein
A.A. Townsend, The structure of turbulent shear flows, 2nd
edn. (Cambridge University Press, Cambridge, 1976)
B. Bamieh, M. Dahleh, Energy Amplification in Channel Flows with Stochastic Excitation, preprint (1999)
(http://online.itp.ucsb.edu/online/hydrot00/bamieh/)
R.F. Blackwelder, L.S. Kovasznay, Phys. Fluids 15, 1545
(1972)
S.B. Pope, Phys. Fluids 14, 1070 (2002)
J. Schumacher, B. Eckhardt, Europhys. Lett. 52, 627
(2000); J. Schumacher, J. Fluid Mech. 441, 109 (2001)
B. Eckhardt, J. Schumacher, A. Jachens, Asymmetric time
correlations in turbulent shear flows. Proceedings IUTAM
meeting on turbulent shear flows, Princeton 2002
O. Cardoso, D. Marteau, P. Tabeling, Phys. Rev. E 49,
454 (1994)
B.S. Williams, D. Marteau, J.P. Gollub, Phys. Fluids 9,
2061 (1997)
U. Brosa, Z. Naturforsch. A 46, 473 (1991)
N. Tillmark, P.H. Alfredsson, J. Fluid Mech. 235, 89
(1992)
F. Daviaud, J. Hegseth, P. Bergé. Phys. Rev. Lett. 69,
2511 (1992)
E.-J. Kim, B. Dubrulle, Phys. Plasmas 8, 813 (2001)
B. Eckhardt, R. Pandit, to be published
Download